Add support for multipoint query retrieval to TFRS Retrieval task.

PiperOrigin-RevId: 689532029

Author: TensorFlow Recommenders Authors

tensorflow_recommenders/tasks/retrieval.py

@@ -141,7 +141,10 @@ def call(
 
     Args:
       query_embeddings: [num_queries, embedding_dim] tensor of query
-        representations.
+        representations, or [num_queries, num_heads, embedding_dim]. If latter,
+        we do "maxsim" scoring over those multiple query heads. This applies to
+        the loss computation and batch metrics. Factorized metrics won't be
+        computed in this case.
       candidate_embeddings: [num_candidates, embedding_dim] tensor of candidate
         representations. Normally, `num_candidates` is the same as
         `num_queries`: there is a positive candidate corresponding for every
@@ -166,8 +169,15 @@ def call(
       loss: Tensor of loss values.
     """
 
-    scores = tf.linalg.matmul(
-        query_embeddings, candidate_embeddings, transpose_b=True)
+    if len(tf.shape(query_embeddings)) == 3:
+      scores = tf.einsum(
+          "qne,ce->qnc", query_embeddings, candidate_embeddings
+      )
+      scores = tf.math.reduce_max(scores, axis=1)
+    else:
+      scores = tf.linalg.matmul(
+          query_embeddings, candidate_embeddings, transpose_b=True
+      )
 
     num_queries = tf.shape(scores)[0]
     num_candidates = tf.shape(scores)[1]
@@ -203,7 +213,7 @@ def call(
     for metric in self._loss_metrics:
       update_ops.append(metric.update_state(loss))
 
-    if compute_metrics:
+    if compute_metrics and len(tf.shape(query_embeddings)) == 2:
       for metric in self._factorized_metrics:
         update_ops.append(
             metric.update_state(

tensorflow_recommenders/tasks/retrieval_test.py

@@ -250,5 +250,53 @@ def test_task(self):
     self.assertAllClose(expected_metrics2, metrics2_)
 
 
+class RetrievalTestWithMultipointQueries(tf.test.TestCase):
+
+  def test_task(self):
+
+    query = tf.constant(
+        [[[3, 2, 1], [1, 2, 3]], [[2, 3, 4], [4, 3, 2]]], dtype=tf.float32
+    )
+    candidate = tf.constant([[0, 1, 0], [0, 1, 1], [1, 1, 0]], dtype=tf.float32)
+    candidate_dataset = tf.data.Dataset.from_tensor_slices(
+        np.array([[0, 0, 0]] * 20, dtype=np.float32)
+    )
+
+    task = retrieval.Retrieval(
+        metrics=metrics.FactorizedTopK(
+            candidates=candidate_dataset.batch(16), ks=[5]
+        ),
+        batch_metrics=[
+            tf.keras.metrics.TopKCategoricalAccuracy(
+                k=1, name="batch_categorical_accuracy_at_1"
+            )
+        ],
+    )
+
+    # Scores will have shape [num_queries, num_candidates]
+    # All_pair_scores:   [[[2,2], [3,5], [5,3]], [[3, 3], [7,5], [5,7]]].
+    # Max-sim scores:    [[2, 5, 5], [3, 7, 7]].
+    # Normalized logits: [[0, 3, 3], [1, 5, 5]].
+    expected_loss = -np.log(1 / (1 + np.exp(3) + np.exp(3))) - np.log(
+        np.exp(5) / (np.exp(1) + np.exp(5) + np.exp(5))
+    )
+
+    expected_metrics = {
+        "factorized_top_k/top_5_categorical_accuracy": (
+            0.0
+        ),  # not computed for multipoint queries
+        "batch_categorical_accuracy_at_1": 0.5,
+    }
+    loss = task(
+        query_embeddings=query,
+        candidate_embeddings=candidate,
+    )
+    metrics_ = {metric.name: metric.result().numpy() for metric in task.metrics}
+
+    self.assertIsNotNone(loss)
+    self.assertAllClose(expected_loss, loss)
+    self.assertAllClose(expected_metrics, metrics_)
+
+
 if __name__ == "__main__":
   tf.test.main()